Two-Level Continuous Monitoring of Work-Related Psychosocial Risk: Formalization and Simulation-Based Proof of Concept of a Sector–Individual Assessment Architecture

Technical Summary: Two-Level Continuous Monitoring of Work-Related Psychosocial Risk

Problem Statement
Regulatory frameworks, specifically Brazil's Regulatory Standard NR-1 (enforcement starting May 26, 2026) and ISO 45003, mandate the continuous assessment and control of work-related psychosocial risks. However, these regulations prescribe what must be achieved (identification, assessment, control) without defining how to measure exposure and response. Existing instruments like the Copenhagen Psychosocial Questionnaire (COPSOQ) and the HSE Management Standards Indicator Tool are ill-suited for this mandate because they are:

1. Cross-sectional and single-informant: They rely on worker self-report for both exposure and outcome, making them vulnerable to common method variance and structurally incapable of detecting "masked risk" (situations where organizational load is high but workers do not report strain due to suppression).
2. Single-level: They measure exposure at the individual level rather than the sector/organizational level, failing to capture hazards inherent to work design.
3. Static: They are administered periodically against normative references, preventing the detection of within-person and within-sector changes required for continuous risk management.

Methodology: The NEXUS Architecture
The paper formalizes and tests a proof-of-concept monitoring architecture called NEXUS (Nested EXposure–response Unified System). The system separates the measurement of exposure (organizational load) from response (individual psychosocial state) and operates on a two-level, longitudinal basis.

• Level 1: Organizational Load Index (OLI) – Sector Level

  • Concept: Quantifies psychosocial load imposed by the organizational arrangement of a sector/function.
  • Domains: Maps to 13 Brazilian Ministry of Labour risk factors, grouped into five domains: Demands (D1), Autonomy/Role Clarity (D2), Relationships/Protection (D3), Recognition/Justice (D4), and Work Organization (D5).
  • Data Source Hierarchy: To ensure independence from worker self-report, the OLI uses a weighted composite of sources ordered by reliability and independence:
    • Primary (ρ=0.60): Objective organizational records (hours, overtime, turnover, absenteeism).
    • Secondary (ρ=0.25): Structured manager checklists (capped to prevent overriding primary records).
    • Tertiary (ρ=0.15): Worker reports on the environment (not self).
  • Calculation: Domain scores are standardized against a "frozen" sector baseline using empirical-Bayes shrinkage to stabilize variance estimates.

• Level 2: Psychosocial Response Index (PRI) – Individual Level

  • Concept: Quantifies individual psychosocial response using a multifrequency battery (PANAS-neg, DASS-21, OLBI, SRQ-20).
  • Calculation: Components are standardized as intra-individual z-scores against the worker's own frozen baseline. This follows an intensive longitudinal tradition where the unit serves as its own reference.

• Level 3: Adherence Telemetry – Non-Verbal Channel

  • Concept: Records behavioral latency, completion rates, and abandonment of the monitoring probes themselves.
  • Architecture: Requires strict production data-engineering specifications (e.g., separating network latency from behavioral latency, immutable audit trails) to ensure these metrics reflect behavioral withdrawal rather than technical artifacts.

• Alert Taxonomy
  Risk is classified at the intersection of OLI and PRI levels using refined criteria (thresholds + persistence rules):

  • Class A (Reactivity): High Load + High Response (Confirmed risk).
  • Class B (Dissonance): High Load + Low/Normal Response (Masked risk; the primary target of the architecture).
  • Class C (Chronicity): Normal Load + High Response (Residual strain).
  • Class D (Disengagement): Behavioral withdrawal (latency/completion issues) without high load or reported strain.
  • Note: A critical protocol for suicidal ideation (SRQ-20) operates independently and immediately.

• Governance
  The system enforces a separation of views: employers see only sector-level aggregates (OLI, risk matrices), while individual signals and flags are visible only to the responsible psychologist, ensuring compliance with data protection laws (LGPD).

Simulation-Based Proof of Concept
The architecture was evaluated via a Monte Carlo study (5 scenarios × 500 replications) simulating 20 workers over 26 biweekly windows.

• Generative Model: Included lagged load-response coupling, verbal suppression (for masked risk), chronic accumulation, and behavioral drift.
• Refinement Process: Initial "naive" criteria (single-window thresholds, raw variance) failed, producing false-alarm rates up to 99% due to multiplicity and baseline noise. The model was refined using:
  1. Empirical-Bayes shrinkage of baseline variance.
  2. Individual-level evaluation.
  3. Persistence requirements (e.g., 3 consecutive windows for Class B).

Key Results
Under the refined criteria, the system demonstrated:

• False Alarm Control: Per-worker annual false-alarm rates ranged from 0.0% to 4.8% under the null scenario.
• Sensitivity:
  • Class A (Reactive): 83.8%
  • Class B (Masked/Dissonant): 72.8% (The configuration invisible to single-informant tools).
  • Class C (Chronic): 93.6%
  • Class D (Disengagement): 95.0%
• Sensitivity Analysis: The system's false-alarm rates were stable across variations in OLI domain weights. However, the detection of load-dependent classes (A and B) was highly sensitive to the weighting of the "Demands" domain. Joint perturbation of weights significantly degraded Class B detection (dropping to 17.9% in some stress tests), indicating that empirical calibration of weights is a prerequisite for field use.

Significance and Claims
The paper claims to establish internal coherence and a formal specification for a monitoring architecture that addresses the structural limitations of current tools.

• Novelty: It is the first formalization of a two-level architecture that separates exposure (sector-level, source-independent) from response (individual-level, longitudinal), specifically designed to detect "masked risk" (Class B).
• Regulatory Alignment: It operationalizes the requirements of NR-1 and ISO 45003 for continuous, documented risk management with an audit trail suitable for inspection.
• Modesty of Claims: The authors explicitly state that the study does not establish validity. The simulation relies on the architecture's own assumptions (e.g., that primary records accurately track load). The parameters (weights, thresholds, reliability coefficients) are declared as initial values pending field calibration. The paper concludes that field validation is the necessary next step before the system can be used for operational decisions.